Components of CNNs (part 1)

In the second half of Brandon's video (Components of CNNs), the main learning objective is: 

  • How strides and pooling operations help us downsample the feature tensors. 
  • What the overall architecture might look like. 
  • How the extracted features can provide information about the class/label of the image. 

 

One misleading detail in this part of Brandon's video: 

  • In classification problems, we normally use a softmax (or a sigmoid) function in the output layer to produce probabilities. To keep things simple, this step is omitted in this video and he ends with what corresponds to the input to the softmax function.